Upload configuration_chest2vec.py with huggingface_hub

configuration_chest2vec.py

@@ -1,10 +1,11 @@
-"""Configuration for Chest2Vec — a LoRA-tuned Qwen3-Embedding model for
-chest radiology report embeddings.
+"""Configuration for Chest2Vec — a Qwen3-Embedding model (LoRA merged in) for chest
+radiology report embeddings.
 
-Chest2Vec = Qwen3-Embedding base + contrastive LoRA adapter. It produces a
-single L2-normalized report embedding (last-token / EOS pooling), matching the
-Qwen3-Embedding convention.
+Self-contained: the merged encoder weights ship in the repo, so loading needs neither the
+`chest2vec` package nor a download of the base Qwen3-Embedding weights. Produces a single
+L2-normalized report embedding (last-token / EOS pooling), Qwen3-Embedding convention.
 """
+from typing import Optional
 from transformers import PretrainedConfig
 
 
@@ -13,14 +14,18 @@ class Chest2VecConfig(PretrainedConfig):
 
     def __init__(
         self,
+        encoder_config: Optional[dict] = None,
         base_model: str = "Qwen/Qwen3-Embedding-0.6B",
-        adapter_subdir: str = "contrastive",
-        require_flash_attention_2: bool = True,
+        hidden_size: int = 1024,
         default_max_len: int = 512,
+        pooling: str = "last_token",
+        matryoshka_dims: Optional[list] = None,
         **kwargs,
     ):
+        super().__init__(**kwargs)
+        self.encoder_config = encoder_config or {}
         self.base_model = base_model
-        self.adapter_subdir = adapter_subdir
-        self.require_flash_attention_2 = require_flash_attention_2
+        self.hidden_size = hidden_size
         self.default_max_len = default_max_len
-        super().__init__(**kwargs)
+        self.pooling = pooling
+        self.matryoshka_dims = matryoshka_dims or []